计算与应用讨论班——Topology-enhanced machine learning for consonant recognition

报告人：朱一飞（南方科技大学）

时间：2025年4月16日，15:30-16:30

地点：海纳苑2幢203室

摘要：In artificial-intelligence-aided signal processing, existing deep learning models often exhibit a black-box structure.  The integration of topological methods serves the dual purpose of extracting structural information from time-dependent data as well as making models more interpretable.  In this talk, I will give an overview of joint work with Pingyao Feng, Qingrui Qu, et al., in which we propose a transparent methodology, TopCap, to capture the most salient topological features inherent in time series for basic machine learning.  Rooted in high-dimensional ambient spaces, TopCap is capable of capturing features rarely detected in datasets with low dimensionality.  Applying time-delay embedding and persistent homology, we obtain descriptors that encapsulate information such as the vibrations of a time series.  This information is then vectorized and fed into multiple machine learning algorithms such as k-nearest neighbors and support vector machines.  Notably, in classifying voiced and voiceless consonants, TopCap achieves an accuracy exceeding 96%, consistently standing in comparison with state-of-the-art neural networks.  Moreover, we integrate TopCap features into those neural networks, beyond direct biomimetic spectral engineering currently adopted in the field.  This has made a network more interpretable with better performance in terms of accuracy, steadiness, convergence of loss function, and robustness against noise.

联系人：蔺宏伟（hwlin@zju.edu.cn）